The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided topics to generate complex HTML and JavaScript slide decks, creating a surface for indirect prompt injection.
Ingestion points: The $ARGUMENTS in SKILL.md (specifically the topic argument) are interpolated into the slide generation process.
Boundary markers: The skill uses a <task> tag in references/create.md, but lacks explicit instructions for the agent to validate or sanitize potentially malicious content within the user-provided topic.
Capability inventory: The skill generates HTML documents in references/html-template.md which include <script> tags and external library integrations, providing a potential execution environment for injected payloads.
Sanitization: No explicit sanitization or escaping logic is defined for the interpolated user content.
[COMMAND_EXECUTION]: The knowledge base files provide specific shell commands for the agent to execute local scripts.
Evidence: Files references/copywriting-formulas.md, references/layout-patterns.md, and references/slide-strategies.md all contain commands like python .claude/skills/design-system/scripts/search-slides.py. This exposes a command execution surface that relies on scripts in a separate skill directory (design-system).
[EXTERNAL_DOWNLOADS]: The skill references external assets from well-known services.
Evidence: references/html-template.md includes a script tag for https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js. This is a well-known and trusted library.
[METADATA_POISONING]: There is an inconsistency between the skill's internal metadata and the platform-provided author information.
Evidence: The SKILL.md frontmatter identifies the author as claudekit, while the platform metadata identifies the author as nextlevelbuilder.

slides